About Horizon · Kira

What this page is, and how Kira uses Horizon differently from Luna.

What Horizon is

Horizon is where you coach Kira by looking at real conversations. You read transcripts, flag replies that weren’t great, update Kira’s Playbook (the metaprompt), and track what you’ve already addressed as named Focus Areas.

The Kira loop, in plain English

Kira’s Horizon is observatory-style: you diagnose by reading real users, not by running synthetic evaluations. (Luna’s Horizon is laboratory-style and has eval runs — see below.)

Step 1·Read

Browse recent Kira sessions in Sessions. Skim for ones that felt off — user dropped, bot looped, didn’t bridge to the assessment, etc. Click in to see the full transcript.

Go to Read →

Step 2·Flag

Click 🚩 on any assistant reply you’d have handled differently. Flagged turns collect in Flagged turns so you can work through them without re-reading whole sessions.

Go to Flag →

Step 3·Diagnose

When a pattern keeps showing up across sessions, name it — create a Focus Area on the Horizon home. Each Focus Area is one behavior you want to coach Kira toward. Tracked as a hypothesis you iterate on.

Go to Diagnose →

Step 4·Edit the Playbook

Open Playbook. This is Kira’s live metaprompt (currently KYF v1.9). Add a rule, tweak a section, save a new version, make it current.

Go to Edit the Playbook →

Step 5·Watch and repeat

Check Activityperiodically to see publishes, reverts, and new flagged turns. Re-read sessions to confirm the pattern you coached out actually stopped appearing. Mark the Focus Area improved when you’re satisfied.

Go to Watch and repeat →

What’s different from Luna

No eval runs.Luna has synthetic eval batches with auto-scoring; Kira doesn’t (yet). Your feedback loop is reading real users, not running reproducible tests.
No Sanity Check / before-after gate. You publish changes and then watch real sessions to see the effect. Cycle time is days, not minutes.
No Guardrails / Safety Rules surface. Kira’s runtime nudges (if any) live in the metaprompt itself, not as separate Horizon triggers.
No Sandbox / Metaprompt Lab. You edit the Playbook directly in the Playbook page. Test in production by observation, not in a sandbox.
Focus Areas are TODOs. Without eval runs the 5-row stage progress on each card is mostly cosmetic — the card is a named hypothesis with a link to Sessions.

Key terms

Everything Horizon calls out in the Kira view, in one place: Open the Glossary →

Where the data lives

Kira sessions, messages, and feedback live in the Kira backend’s PostgreSQL database. The Horizon frontend talks to it through two routes: /dashboard/api/* for session-inspector data, and /api/horizon/* for Focus Areas, Activity, and Playbook events (shared with Luna on the same shared-backend server, filtered by product_id=kira).

Keyboard shortcuts

? — open this page’s tour
g h — go to Horizon home
g g — go to Glossary
g a — go to Activity
g p — go to Playbook
Esc — close tour / modal